Dynamic Deep Coding for Secure Image Transmission Over Communication Networks

Deep learning-based joint source-channel coding is an efficient method for wireless image transmission over additive white Gaussian noise (AWGN) channels. It enjoys high data compressibility and insensitivity to noise, especially at low signal-to-noise ratios. Current approaches are unable to adjust code rates based on diverse channel conditions and image contents and hence become inflexible and bandwidth-inefficient. To solve this issue, variable-length deep learning modeling offers variable code length adaptability, and an oracle network predicts the peak signal-tonoise ratio (PSNR) from image content, channel SNR, and compression ratio. A compression optimizer then compresses data transfer while keeping images at a set quality. These methodologies promote bandwidth efficiency by sending images with the smallest possible data size but keeping quality high. Simulation results show that the proposed approach generates image quality on par with state-of-the-art attention models, and the Oracle network predicts PSNR values within less than 0.5 dB average deviation. The approach is capable of efficiently reducing bandwidth usage but can also cope with different channel conditions and classes of images, and it is a reliable option for secure and efficient image transmission in AWGN networks.